The Double-Edged Nature of the Rashomon Set for Trustworthy Machine Learning